CT Scans May Predict Which Brain Injury Patients Will Need a Second Surgery

After emergency surgery to drain a blood clot from the brain, some traumatic brain injury patients stabilize. Others deteriorate. Their brains swell, intracranial pressure climbs, and surgeons must perform a second, more aggressive operation called decompressive craniectomy, in which a section of skull is removed to give the swelling brain room to expand. The question clinicians desperately want answered is: which patients are heading for that second surgery?

A study from Central South University in China, published in January 2026 in the Chinese Neurosurgical Journal, suggests the answer may already be embedded in the CT scans taken before the first operation. The research team, led by Dr. Zhongyi Sun, used a technique called radiomics to extract subtle imaging features invisible to the human eye and trained machine learning algorithms to predict which patients would need secondary decompressive craniectomy.

Reading what radiologists cannot see

Radiomics converts medical images into quantitative data. Instead of a radiologist interpreting a CT scan as showing "significant edema" or "midline shift," a radiomics algorithm measures dozens of numerical features describing tissue shape, texture, and intensity patterns. These features can capture pathological changes too subtle or complex for visual assessment.

The team analyzed pre-evacuation CT scans from 65 adult patients who had undergone emergent craniotomy for hematoma evacuation and bone flap replacement after traumatic brain injury. Some of these patients later required secondary decompressive craniectomy due to refractory intracranial hypertension. The researchers extracted more than 100 radiomic features from each scan and fed them into several machine learning models.

Imaging features versus clinical indicators

The comparison was telling. Models built on demographic and clinical information alone, the kind of data clinicians typically use for risk assessment, performed poorly. They could not reliably distinguish between patients who would and would not need a second surgery.

Models using radiomic features from the CT scans performed substantially better, accurately identifying high-risk patients. And when imaging features were combined with selected clinical variables, prediction improved further. The best-performing approach used both data types together, suggesting that radiomics complements rather than replaces clinical judgment.

From reactive treatment to early warning

The clinical value of early prediction in this context is significant. Decompressive craniectomy is a major operation with substantial risks. Patients who need it often deteriorate rapidly, leaving neurosurgeons with narrow decision windows. If a model could flag high-risk patients within hours of their first surgery, clinicians could increase monitoring intensity, prepare surgical teams, and potentially intervene before intracranial pressure reaches dangerous levels.

The approach could also influence resource allocation. Intensive care beds, neurosurgical operating room time, and specialized monitoring equipment are all limited. Knowing which patients are most likely to need them allows hospitals to plan rather than react.

Significant caveats

The study has several important limitations. The sample of 65 patients is small for a machine learning study. Models trained on small datasets risk overfitting, meaning they may perform well on the training data but fail to generalize to new patients at different hospitals with different scanners and imaging protocols.

The study was conducted at a single center, and no external validation on an independent patient cohort was reported. Until the model is tested on patients from multiple institutions, its real-world reliability remains uncertain.

The radiomic features were extracted from pre-surgery CT scans. Whether post-operative imaging, clinical trajectory data, or intracranial pressure measurements could improve predictions was not explored. A more comprehensive model incorporating multiple data streams might perform differently.

Traumatic brain injury is also highly heterogeneous. The mechanisms, severity, location, and associated injuries vary enormously between patients. A model trained on 65 cases from one institution may not capture that variability adequately.

The team acknowledges these gaps and hopes that future multicenter studies with larger patient populations and automated imaging workflows will refine the model and move it closer to clinical use.

The broader direction

This study fits into a growing body of research exploring whether routine medical imaging contains predictive information that standard visual interpretation misses. Radiomics has shown promise in oncology, cardiology, and pulmonology. Its application to traumatic brain injury, where rapid clinical deterioration can be fatal, represents a logical extension.

If validated at scale, such tools could change the rhythm of neurointensive care, shifting the focus from crisis response to anticipatory management. For a condition that disproportionately affects young people and carries lifelong consequences, that shift matters.